acceptodds
Under review as a conference paper at ICLR 2027

Dirac Convolutional Neural Networks for Simplicial Complexes

24 views

Abstract

Higher-order systems capture interactions that go beyond the pairwise relationships represented by graphs and arise naturally in models of biological and physical systems. Simplicial complexes provide a common framework for representing such higher-order interactions, motivating the development of neural architectures that operate on them. In this work, we introduce Dirac Convolutional Networks (DCNs), a family of simplicial neural networks (SNNs) based on Dirac convolutions we introduce. Our construction is derived from a spectral filter formulation for multichannel, graded signals on simplicial complexes. We instantiate our framework with normalized root-to-leaf Dirac operator and propose Root-to-Leaf DCNs (R2L-DCNs) that inherit the benefits of normalized operators. First, normalization fixes the spectral scale of the Dirac operator and Hodge Laplacians, allowing filters to transfer across simplicial complexes. Second, it improves the stability of convolutional SNNs, allowing deep neural architectures. We establish this by proving that continuous, linear versions of these networks can preserve both the Dirichlet energy and the norm with right weights, and that their Euler discretizations are more stable when normalized operators are used. Finally, we empirically validate these properties across multiple benchmarks, showing that R2L-DCNs outperform existing SNN architectures on trajectory prediction and mesh classification benchmarks, and can maintain performance as network depth increases.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Simplicial Complex Neural Networks, Graph Neural Networks, Spectral Filters

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cnweOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dirac Convolutional Networks (DCNs), a family of simplicial neural networks built from polynomial spectral filters of the Dirac operator, which recovers existing convolutional SNNs such as SCCNN and SCNN as special cases. The authors instantiate DCNs with the root-to-leaf normalized Dirac operator of Viganò et al. (2026) and Chebyshev polynomials of order two (R2L-DCNs). They show that a continuous linear version of the model with skew-symmetric weights preserves the norm and the Dirichlet energy, and that the forward-Euler discretization converges to the highest eigenvalue of the Hodge Laplacian, with rates that depend on the spectrum. The method is evaluated on trajectory prediction (ocean drifters), mesh classification (SHREC'11), node and graph classification, and a depth experiment on a rewired version of minesweeper.

Critical strengths

The paper is easy to follow, and the theoretical results and their proofs are clearly presented. Using a normalization that bounds the spectrum of the Dirac operator and all Hodge Laplacians in a common interval is a clean idea, and it makes Chebyshev filters natural in this setting. R2L-DCN obtains competitive results with the baselines, and Figure 4 shows that it keeps its performance when stacking many layers, whereas SCCNN degrades.

Critical weaknesses

I recommend rejection in its current form. The idea is clean, but the contribution is limited with respect to existing simplicial convolutional networks, and the main claims (transferability and stability) are not supported by the experiments or the theory.

  • Limited novelty. Even powers of the Dirac operator are powers of the Hodge Laplacians, and odd powers are (co)boundary maps. With order two, R2L-DCN reduces to identity, up/down Hodge Laplacian, and incidence terms, i.e., an SCCNN layer with normalized operators. The main difference with SCCNN is therefore the normalization, which comes from Viganò et al. (2026). Moreover, Dirac-based polynomial filters for simplicial neural networks were already proposed in [R1], which isn't cited or discussed. The ablations also don't show a clear benefit of the inter-rank coupling: R2L-DCN(even), which removes it, is better than R2L-DCN on both irregular meshes in Table 1 (48.0 vs. 46.8 and 45.4 vs. 41.1).
  • The main claims are not backed by experiments or theory.
    • The abstract and the contributions claim that normalization allows filters to transfer across simplicial complexes, but there is no transferability experiment, and the conclusion lists "testing transfer to unseen complexes" as future work. Normalization guarantees that the spectra of different complexes lie in the same interval, but this isn't enough to claim that learned filters transfer: the informative eigenspaces can still be at different frequencies. The book-complex example in Figure 2 is a hand-picked illustration, not evidence.
    • The theory doesn't directly apply to the proposed model. The analysis considers linear dynamics with only the first-order term and constrained (symmetric or skew-symmetric) weights, while R2L-DCN uses nonlinearities, unconstrained weights, and second-order Chebyshev terms. Moreover, Proposition 4.1 holds for any symmetric propagation operator with skew-symmetric weights, so it's unclear what is specific to the Dirac formulation.
    • For the forward-Euler discretization with skew-symmetric weights, Eqn. (38) shows that every mode is amplified by a factor larger than one, so the norm grows with depth. Normalization only slows down this growth, which is weaker than the stability claimed in the abstract.
    • There is no comparison between normalized and unnormalized DCNs with the same architecture, so the effect of R2L normalization on accuracy and depth robustness isn't isolated.

Minor: In Eqns. (16) and (40), the skew-symmetric rate uses θ2\theta^2, where θ\theta is defined as an eigenvalue, which is purely imaginary in this case. It should be the squared modulus, as in Eqn. (38).


References

[R1] "Generalized simplicial attention neural networks", IEEE Transactions on Signal and Information Processing over Networks, 2024.

Questions for authors

Clarifications:

  • Is R2L-DCN of order two equivalent to an SCCNN with R2L-normalized operators? If not, what can it represent that SCCNN can't? How is DCN positioned with respect to [R1]?
  • To what extent do the theoretical results explain the behavior of R2L-DCN, given that the analysis focuses on linear dynamics with constrained weights? What is specific to the Dirac formulation in Proposition 4.1?
  • In what precise sense is the Euler discretization stable, given that the norm grows with depth for skew-symmetric weights?

Suggestions for the revision:

  • Provide experiments showing the transferability of R2L-DCNs across simplicial complexes, e.g., training on some complexes and testing on unseen ones with different sizes or degree distributions.
  • Compare normalized and unnormalized DCNs with the same architecture to isolate the effect of R2L normalization on accuracy and depth robustness.
8 comments
AuthorsOpenReview ↗

Official Comment by Authors - Part 1:

We sincerely thank the reviewer for their detailed feedback and constructive suggestions. We are especially encouraged that the reviewer views our contributions as clean and natural, and that the significance of retaining performance with depth is acknowledged by the reviewer. We address the concerns below:

Novelty

Limited novelty. Even powers of the... // Is R2L-DCN of order two equivalent to an SCCNN... If not, what can it represent that SCCNN can't? How is DCN positioned with respect to [R1]?

We appreciate the reviewer for bringing [R1] to our attention, and we have cited it in the revised work. We point out the main differences below:

  • Formulating in spectral domain: Our formulation of Dirac convolutions fundamentally differs from the presentation in [R1], as we define it in the spectral domain with a linear map acting on each eigenspace, which then yields the polynomial formulation as derived in the proof of proposition 3.4. Our formulation in the spectral domain, with multi-channel signals, requires careful treatment leveraging chirality.
  • Parametrizing the spectral filter: In order to show which polynomial bases can provide the 'block-diagonal'/'block-anti-diagonal' parametrization, we provide the polynomial basis to satisfy Tk(−x)=(−1)kTk(x)T_k(-x)=(-1)^kT_k(x), which allows the use of Chebyshev polynomials, and it is a non-trivial fact to show. For an arbitrary polynomial basis, such a block-diagonal / block-anti-diagonal parametrization may not correspond to a graded function. This is a detail we enforce in the proof of proposition 3.4.

We believe developing new spectral learning methods on simplicial complexes is a crucial contribution, and our formulation can guide next steps such as non-linear filters which respect the coupling on simplicial complexes. Our work also makes it explicit that different kinds of signals may lie on different ranks, and what nuances a graded domain brings when we develop spectral methods.

We agree that the R2L-DCN architecture can be viewed as a replacement of the unnormalized operators with their R2L-normalized counterparts in SCCNN, but we also believe that this is a reductive view of our contributions. The main contribution of the paper is not the R2L-DCN architecture, it is only an instantiation of our framework, which explores:

  • How one can directly construct convolution directly using the graded structure of simplicial complexes
  • The flexibility of the choice in the Dirac operator, which can be normalized, unnormalized, weighted, or random-walk Dirac operators
  • Which polynomial bases create an interpretable parametrization, where odd powers correspond to inter-rank relations, and even powers correspond to intra-rank relations

With this perspective, [R1] and SCCNNs become a special case of our broad framework. We hope one can appreciate the significance of such a framework to compare SNNs in a systematic manner, to see what choices are useful in which tasks. In our new filter transferability experiments, we also use the random-walk Dirac operators which fit into our framework with a simple pre-/post-processing step, showing the flexibility of our framing.

Furthermore, a significant part of our contribution is the theoretical analysis of the asymptotic behaviour of the Dirichlet energy and the Rayleigh quotient with the linear Dirac evolution, and characterizing its possible limits with symmetric and skew-symmetric weights. To the best of our knowledge, we provide first such analysis where the Dirac operator interacts with the weight matrices acting on the features. Our result shows that with (skew-)symmetric weight matrices, regardless of the learned weights, the Dirac operator tends to act a a high-pass filter. This result is novel compared to previous analysis on coupled SNNs and analogous results in GNNs, and exploring coupled dynamics is crucial for extending notions of oversmoothing to simplicial complexes.

Finally, we also discuss the interpretation of normalization from a spectral-scale perspective, which has not been made clear in the SNN case. We believe that such discussions will be helpful to decide which architecture should be used for certain tasks.

We hope to have conveyed that our work is indeed novel, goes beyond [R1] and SCCNNs in our spectral framing, theoretical analysis, and practical implications. We believe that our exposition, now strengthened by this discussion is crucial to understand the use cases of topological deep learning by emphasizing the significance of the choices we make while designing the architectures.

The ablations also don't show a clear benefit of the inter-rank coupling...

We want to clarify that, our argument in the paper is not that inter-rank coupling is always beneficial. We introduce a framework which encompasses the possible choices that can be made for SNN architectures, informing a broad, task-dependent design of simplicial complex networks.

AuthorsOpenReview ↗

Official Comment by Authors - Part 2:

Filter Transferability

The abstract and the contributions claim that normalization allows filters to transfer across simplicial complexes, but there is no transferability experiment,

We appreciate this comment, and now conduct experiments to address reviewer's concern. In particular, we design an edge-classification task, where the goal is to detect which edges are corrupted via ±1\pm1 in the existence of a gradient signal coming from the vertices, where the filters are trained on a small (60 vertices) simplicial complex and transferred to larger complexes. We added Section 5.3 and Appendix D.2 for further details and discussion of the problem. We also discuss the random-walk variant of R2L upon the suggestion of another reviewer.

Table A: Balanced accuracy (%) in corrupted-edge detection task for filters of order K.

OperatorKn=60n=240Irreg. n=5200
Unnorm.189.1 ± 1.982.2 ± 0.277.5 ± 0.0
395.4 ± 0.859.8 ± 0.768.2 ± 0.4
1/μ_\max188.8 ± 1.998.9 ± 0.165.4 ± 0.2
395.8 ± 0.799.9 ± 0.081.5 ± 1.3
R2L187.0 ± 1.296.6 ± 0.196.5 ± 0.0
388.4 ± 1.095.4 ± 0.295.9 ± 0.0
R2L(rw)194.6 ± 0.699.9 ± 0.099.7 ± 0.0
395.9 ± 0.799.9 ± 0.099.8 ± 0.0

Table B: False positive rate (FPR) and false negative rate (FNR) for the filters. Lower is better.

OperatorKn=240 FPRn=240 FNRIrreg. n=5200 FPRIrreg. n=5200 FNR
Unnorm.135.7 ± 0.40.0 ± 0.044.9 ± 0.10.2 ± 0.0
379.5 ± 1.31.0 ± 0.063.1 ± 0.70.6 ± 0.0
1/μ_\max11.1 ± 0.11.2 ± 0.30.3 ± 0.168.8 ± 0.3
30.3 ± 0.00.0 ± 0.00.0 ± 0.037.0 ± 2.7
R2L12.3 ± 0.14.6 ± 0.22.6 ± 0.04.4 ± 0.0
38.5 ± 0.30.8 ± 0.07.2 ± 0.01.1 ± 0.0
R2L(rw)10.1 ± 0.00.1 ± 0.00.2 ± 0.00.4 ± 0.0
30.2 ± 0.00.0 ± 0.00.3 ± 0.00.0 ± 0.0

We see that R2L normalization outperforms the unnormalized filters for transferability in this task. This happens as an unnormalized filter extrapolates the frequency response to unseen eigenvalues, where it disproportionately scales up the outputs compared to the learned threshold. R2L also outperforms the 1/\mu_\max normalization when the underlying simplicial complex is irregular. This is because the irregularity is caused by a dense, local cluster, and 1/\mu_\max disproportionately shrinks the filter outputs, causing the outputs to remain below the threshold. This is visible in Table B. The performance of R2L(rw) comes from the fact that it changes the basis of the signal to be compatible with the edge-signals, which is discussed in more detail in Appendix D.2.

Normalization guarantees that the spectra of di"erent complexes lie in the same interval, but this isn't enough to claim that learned filters transfer: the informative eigenspaces can still be at di"erent frequencies. The book-complex example in Figure 2 is a hand-picked illustration, not evidence.

We agree with the reviewer's comment that the informative eigenspaces might be at the different components, and it is a strong assumption to make that the informative frequencies remain identical for every frequency. Our argument is not that normalized operators always transfer, but that a shared spectral scale is important if transferability is desired.

When a filter is learned on a fixed domain, we only care about its action on the eigenvalues of the domain. When it is transferred to another domain with different eigenvalues, the polynomial will be acting extrapolating to unseen frequencies, which can have negative impact on the performance, and this view is now demonstrated in Figure 5, Section 5.3. Especially since a polynomial filter will be strictly increasing or decreasing, when the spectral scales differ too much, this can have drastic impact on the scale of the output. Normalization mitigates this, also shown by our new experiments. Book complexes are a theoretically tractable example to demonstrate our point, and hopefully the experiments now complement the picture.

We thank the reviewer for pointing this out, and we have changed our wording in the revision, so that our claim is clearer.

AuthorsOpenReview ↗

Official Comment by Authors - Part 3:

Linear Models, Skew-Symmetry

The theory doesn't directly apply to the proposed model. The analysis considers linear dynamics with only the first-order term and constrained (symmetric or skew-symmetric) weights, while R2L-DCN uses nonlinearities, unconstrained weights, and second-order Chebyshev terms

We like to emphasize that, in this work, we proposed DCNs as a general learning framework, providing both theoretical and empirical contributions. Similar to many previous works including di Giovanni et al., Deep Linear Networks [R2], our theoretical framework considers linear networks and leaves the analysis of networks with non-linear activations for a future study. This is the setting in the Prop. 4.1/4.2. Our empirical analysis on the other hand (e.g. Tab. 1, 2 & 3), deploys non-linear DCNs with normalized operators to practical problems. We kindly ask the reviewer to consider our contributions (DCN framework, linear theoretical analysis and empirical non-linear DCNs) as a whole when making their assessment.

What is specific to Dirac?

Proposition 4.1 holds for any symmetric propagation operator with skew-symmetric weights, so it's unclear what is specific to the Dirac formulation // What is specific to the Dirac formulation in Proposition 4.1?

We agree with the reviewer that any symmetric propagator with skew-symmetric weight matrix gives energy preservation in GNNs, which is explored in Gravina et al. We do not claim the use of skew-symmetry is novel and unique, but the origins of the skew-symmetry have different implications in coupled dynamics.

For a skew-symmetric weight matrix, the evolution of the linear term can be equivalently written as:

(0BiBi⊤0)X[i](0Wi−Wi⊤0)=(0−BiBi⊤0)X[i](0WiWi⊤0)\begin{pmatrix} 0 & \mathbf{B}_{i} \\\\ \mathbf{B}_i^\top & 0 \end{pmatrix}\mathbf{X_{[i]}}\begin{pmatrix} 0 & \mathbf{W}_{i} \\\\ -\mathbf{W}_i^\top & 0 \end{pmatrix} = \begin{pmatrix} 0 & -\mathbf{B}_{i} \\\\ \mathbf{B}_i^\top & 0 \end{pmatrix}\mathbf{X_{[i]}}\begin{pmatrix} 0 & \mathbf{W}_{i} \\\\ \mathbf{W}_i^\top & 0 \end{pmatrix}

which now acts on features with a symmetric weight matrix. Such construction differs from the previous cases in the following ways:

  • Skew-symmetry comes from the structure. It is the bi-partite structure of vertices and edges that create the skew-symmetry and not the skew-symmetry of the weights exactly. Now the structure is described by a skew-symmetric analogue of the Dirac operator (call K[i]\mathbf{K}_{[i]} ). Furthermore, the linear dynamics generated by K[i]\mathbf{K}_{[i]}, that is X˙[i]=K[i]X[i]\dot{\mathbf{X}}_{[i]}=\mathbf{K}_{[i]}\mathbf{X}_{[i]}, can be differentiated again to get X¨[i]=K[i]2X[i]=−L[i]X[i]\ddot{\mathbf{X}}_{[i]}=\mathbf{K}_{[i]}^2\mathbf{X}_{[i]}=-\mathbf{L}_{[i]}\mathbf{X}_{[i]}, that is the rank-wise wave equation. Therefore intuitively, this can be viewed as a first-order way to describe the wave equation. This is the key difference to having skew-symmetric weights. The relationship between how this affects the SNN dynamics is left for future work.
  • Dirac operator enforces different structure on Wi\mathbf{W}_i. For rank-wise interactions with a 1-channel signals, the only way we can act with a skew-symmetric matrix is if W=0\mathbf{W}=\mathbf{0}, which cannot learn anything. Similarly for a dd-channel signal, we enforce the diagonals to be 00, meaning a feature update cannot be informed by the feature itself. With the Dirac operator, Wi\mathbf{W}_i need not satisfy any structure, it can be any matrix. The structure we enforce is that the vertex →\rightarrow edge map is the transpose (times −1-1) of the edge →\rightarrow vertex. This is a different structure enforced, special to the Dirac case.
  • Graded Signal. The vertex and edge signal spaces are assumed to be different kinds, as they will hold different types of information (scalar, flow, circulation...). This differs the Dirac case from a bi-partite graph. In a bi-partite graph, often the signal spaces of both parts are the same, therefore the signal matrix does not necessarily have a block-diagonal structure.

These three points make the skew-symmetry with the Dirac operator fundamentally different than skew-symmetric weight matrices in any rank-wise propagation, and we believe that these create many interesting research directions on the properties of coupled dynamics. We have added this discussion in Appendix E.

Reviewer_U7sZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Dirac Convolutional Networks (DCNs), which split the Dirac operator of a simplicial complex into rank-coupling blocks D[i] and apply a separate learned polynomial filter to each. A parity argument shows that even powers act within ranks and odd powers act across ranks. The instantiated model, R2L-DCN, uses the root-to-leaf normalized Dirac operator with a degree-2 Chebyshev filter. The authors argue that normalization helps filters transfer across complexes and keeps deep models stable, and support this with a linear analysis of continuous-time and forward-Euler dynamics. Experiments cover trajectory prediction, mesh classification, node and graph classification, and a depth study.

Critical strengths
  • The framework places SCCNN, SCNN and the SNN of Ebli et al. in one formulation. This makes the design choices (operator, polynomial basis, coupling) explicit.
  • Bringing R2L normalization into SNNs is a sensible idea. It preserves the chain-complex property (BkBk+1=0)(B_k B_{k+1} = 0) , bounds the spectra of the Dirac operator and all Hodge Laplacians uniformly, and reduces to degree normalization on graphs. The book-complex example in Section 3.3 makes the transfer argument concrete.
  • The proofs are short. As far as I checked (Proposition 3.4, Corollary 3.5, Lemma C.1, Propositions 4.1 and 4.2), they are correct.
  • The experiments cover several task types. They include an uncoupled variant, R2L-DCN(even), and the paper describes how the irregular meshes were generated.
Critical weaknesses
  • The experiments do not isolate normalization. R2L-DCN differs from SCCNN in both the operator and the polynomial basis, and no ablation fixes the basis to compare R2L with constant rescaling or random-walk normalization, even though Section 4.2 concedes that constant rescaling has the same effect on convergence.
  • Most gains are within noise. R2L-DCN beats the strongest competitor by more than one standard deviation on only 2 of 12 datasets and trails on 3.
  • The two practical claims are not supported. Transfer across complexes is never tested, and the depth study uses linear, weight-shared models on minesweeper, a local task where no model gains from depth.
  • The novelty over SCCNN is incremental: R2L-DCN mainly changes the basis and adopts the operator of Viganò et al. (2026). The paper also does not cite SSN [1], which already normalizes boundary, coboundary, up and down operators in a block operator over all ranks.

[1] Madhu, Hiren, Sravanthi Gurugubelli, and Sundeep Prabhakar Chepuri. "Unsupervised parameter-free simplicial representation learning with scattering transforms." Forty-first International Conference on Machine Learning. 2024.

Questions for authors

Please see weaknesses.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors - Part 1

We appreciate the reviewer's detailed review and suggestions. We are grateful that the reviewer finds our framework useful for making the design choices in SNNs explicit, and regards the importance of using R2L normalization. We address each point below:

Novelty:

The novelty over SCCNN is incremental: R2L-DCN mainly changes the basis and adopts the operator of Viganò et al. (2026). The paper also does not cite SSN [1], which already normalizes boundary, coboundary, up and down operators in a block operator over all ranks.

We thank the reviewer for bringing up SSN, we have cited it in the revision of our work. We want to point out the main difference between the normalization in SSN, and root-to-leaf normalization:

Given a signal on a rank, xk∈R∣Xk∣x_k \in \mathbb{R}^{|\mathcal{X_k}|}, Hodge Decomposition allows the signal to be decomposed into 3 orthogonal subspaces, R∣Xk∣≅im Lkd⊕ker⁡ (Lkd+Lku)⊕im Lku\mathbb{R}^{|\mathcal{X_k}|} \cong \text{im}\,L_k^d \oplus \ker\, (L_k^d + L_k^u) \oplus \text{im}\, L_k^u, which follows from the fact that LkuLkd=LkdLku=0L_k^uL_k^d=L_k^dL_k^u=0. The main contribution of the normalization in Viganò et al. (2026) is that the normalized up and down Laplacians still multiply to zero, because of how the normalization is created. In a 'degree-based' normalization (as in SSN), this property no longer holds, hence we do not have a normalized Hodge decomposition. For any topological signal, Hodge decomposition allows a strong analysis tool for interpretation and analysis of models and signals, therefore it is helpful. Our work remains novel in this sense.

While we agree with the reviewer that our network bears similarities with SCCNNs, we kindly like to disagree on the novelty aspect. The use of (R2L) normalization is a crucial step that enables us to build a novel Dirac Convolution framework, its Chebyshev treatment and spectral filters. As such, our paper is a testament to the fundamental nature of Laplacian normalization. In what follows, we highlight the particular distinctions of our framework from the prior works:

  1. Our paper introduces and defines Dirac convolutions in the spectral domain, where filters are frequency responses per eigenvalue of the Dirac operator. Previous work either does this via polynomials of the Dirac operator, or do not discuss how the action of (co)boundary operators can be interpreted in the frequency domain.
  2. In proposition 3.4, we characterize how spectral filters can be parametrized, and provide a criteria on the choice of bases so that the weight matrices are block-diagonal and block-anti-diagonal parameters in even and odd powers, respectively. In SCCNN, the monomial basis is used, and it is non-trivial to show that Chebyshev basis can be used with the same parametrization. This shows that in simplicial complexes, one needs to pay attention on how the coupling plays a role when parametrizing spectral filters, which has not been shown explicitly with the Dirac operator on Simplicial complexes. We believe this is a crucial contribution to develop new spectral methods on simplicial complexes, and this formulation can guide next steps such as non-linear filters which respect the coupling on simplicial complexes.
  3. To the best of our knowledge, we provide the first characterization for how coupled linear dynamics affect asymptotic behaviour of the Rayleigh quotient on simplicial complex neural networks. Our result shows that with (skew-)symmetric weight matrices, regardless of the learned weights, the Dirac operator tends to act a a high-pass filter. This results differs from previous analysis on coupled SNNs and analogous results in GNNs, and exploring coupled dynamics is crucial for extending notions of oversmoothing to simplicial complexes.
  4. Our proposed Dirac Convolutional Networks framework allows a principled way of constructing new simplicial architectures inspired by spectral methods. One could use alternative Dirac operators (e.g. weighted according to certain modeling choices, random-walk Dirac operators) and use different polynomial bases to explore different architectures. Conceptually, many SCNN differ by specific modifications (exclusion of (co)boundary, additional identity term, combination of up and down-weights), and DCNs unify the proposed architectures. Furthermore, we discuss when normalization can be helpful from frequency and domain-irregularity perspectives, which has not been done explicitly in SNNs.

With these in mind, we believe that our contributions are indeed novel, and not incremental, going beyond SCCNN, opening many interesting theoretical and practical research directions in developing spectral methods on SCs that respect the inter-rank coupling, and coupled dynamics. We have clarified our contributions and discussed our work in relation to SSNs and SCCNNs in Appendix D.3.

AuthorsOpenReview ↗

Official Comment by Authors - Part 2:

Filter Transferability

The two practical claims are not supported. Transfer across complexes is never tested...

We appreciate this constructive suggestion. To answer reviewer's concern, we now design an edge-classification task, where the goal is to detect which edges are corrupted via ±1\pm1 in the existence of a gradient signal coming from the vertices, where the filters are trained on a small (60 vertices) simplicial complex and transferred to larger complexes. We added Section 5.3 and Appendix D.2 for further details and discussion of the problem. We also discuss the random-walk (rw) variant of R2L upon the suggestion of the reviewer.

Table A: Balanced accuracy (%) in corrupted-edge detection task for filters of order K.

OperatorKn=60n=240Irreg. n=5200
Unnorm.189.1 ± 1.982.2 ± 0.277.5 ± 0.0
395.4 ± 0.859.8 ± 0.768.2 ± 0.4
1/μ_max188.8 ± 1.998.9 ± 0.165.4 ± 0.2
395.8 ± 0.799.9 ± 0.081.5 ± 1.3
R2L187.0 ± 1.296.6 ± 0.196.5 ± 0.0
388.4 ± 1.095.4 ± 0.295.9 ± 0.0
R2L(rw)194.6 ± 0.699.9 ± 0.099.7 ± 0.0
395.9 ± 0.799.9 ± 0.099.8 ± 0.0

Table B: False positive rate (FPR) and false negative rate (FNR) for the filters. Lower is better.

OperatorKn=240 FPRn=240 FNRIrreg. n=5200 FPRIrreg. n=5200 FNR
Unnorm.135.7 ± 0.40.0 ± 0.044.9 ± 0.10.2 ± 0.0
379.5 ± 1.31.0 ± 0.063.1 ± 0.70.6 ± 0.0
1/μ_\max11.1 ± 0.11.2 ± 0.30.3 ± 0.168.8 ± 0.3
30.3 ± 0.00.0 ± 0.00.0 ± 0.037.0 ± 2.7
R2L12.3 ± 0.14.6 ± 0.22.6 ± 0.04.4 ± 0.0
38.5 ± 0.30.8 ± 0.07.2 ± 0.01.1 ± 0.0
R2L(rw)10.1 ± 0.00.1 ± 0.00.2 ± 0.00.4 ± 0.0
30.2 ± 0.00.0 ± 0.00.3 ± 0.00.0 ± 0.0

We see that R2L normalization outperforms the unnormalized filters for transferability in this task. This happens as an unnormalized filter extrapolates the frequency response to unseen eigenvalues, where it disproportionately scales up the outputs compared to the learned threshold. R2L also outperforms the 1/\mu_\max normalization when the underlying simplicial complex is irregular. This is because the irregularity is caused by a dense, local cluster, and 1/\mu_\max disproportionately shrinks the filter outputs, causing the outputs to remain below the threshold. This is visible in Table B. The performance of R2L(rw) comes from the fact that it changes the basis of the signal to be compatible with the edge-signals, which is discussed in more detail in Appendix D.2.

Depth Study

the depth study uses linear, weight-shared models on minesweeper, a local task where no model gains from depth

We would like to clarify that our main claim in section 5.2 is that R2L-DCN does not lose performance from depth, while SCCNN, which uses an unnormalized operator, does lose performance. We agree with the reviewer that minesweeper is a local task with no gains from the depth, but our point is that even in such a local task, an unnormalized model loses performance. In our observation, this happens as the model needs to decrease the weight matrix norm in order to balance the higher eigenvalues of the unnormalized operator, while the normalized model does not seem to suffer from it in our example. In future applications of SNNs where a model gains from depth, it is crucial that the model does not converge due to geometric bottlenecks, and our point is that normalization is an important step in this process.

AuthorsOpenReview ↗

Official Comment by Authors - Part 3:

Ablations

The experiments do not isolate normalization. R2L-DCN di"ers from SCCNN in both the operator and the polynomial basis, and no ablation fixes the basis to compare R2L with constant rescaling or random-walk normalization, even though Section 4.2 concedes that constant rescaling has the same e"ect on convergence.

We thank the reviewer for mentioning further comparisons to isolate normalization. In this direction, we add ablations on the ocean drifters dataset with 1/μmax⁡1/\mu_{\max}, 1/μmax⁡1/\mu_{\max}-Cheb with the Chebyshev polynomials, R2L-rw using the random walk Laplacian, R2L-Monomial with the monomial basis.

Methodoceanocean (+150)ocean (+600)
R2L-Monomial57.7 ± 5.751.1 ± 9.845.5 ± 4.6
1/μmax⁡\mu_{\max}51.5 ± 6.950.1 ± 4.842.3 ± 6.5
1/μmax⁡\mu_{\max}-Cheb53.9 ± 5.048.8 ± 6.948.7 ± 5.6
R2L-DCN56.7 ± 5.346.8 ± 8.341.1 ± 4.0
R2L-Monomial (even)50.8 ± 5.441.6 ± 3.836.4 ± 4.5
R2L-rw (even)52.7 ± 4.543.9 ± 5.140.1 ± 6.6
1/μmax⁡\mu_{\max} (even)48.9 ± 7.245.1 ± 5.235.8 ± 8.7
1/μmax⁡\mu_{\max}-Cheb (even)53.6 ± 5.645.6 ± 7.343.4 ± 3.9
R2L-DCN (even)53.2 ± 5.448.0 ± 6.045.4 ± 5.0

In coupled models, R2L-Monomial beats all models in ocean and ocean(+150), performing reasonbly well in ocean(+600), where 1/μmax⁡\mu_{\max}-Cheb excels. In rank-wise models and R2L-DCN (even) performs the best in ocean(+150) and ocean(+600), and 1/μmax⁡\mu_{\max}-Cheb (even) performs best in ocean. Referring to the discussion in the previous part, the similar performance of the normalized models can be explained through the fact that the task is highly dependent on the harmonic frequency.

Most gains are within noise. R2L-DCN beats the strongest competitor by more than one standard deviation on only 2 of 12 datasets and trails on 3.

We hope that the experiments on filter transferability gave further evidence on the usefulness of R2L normalization compared to other methods. In ocean drifters dataset, the frequency of interest is 0, i.e. the harmonic component, which unnormalized operators keep consistent across complexes. We do not claim that normalization is a process that makes models perform better in every task, but there are cases where it is beneficial, and the main contribution of our paper is to provide the framework so different SNNs can be compared through the same lens.

We also would like to clarify that R2L-DCN is an instantiation of our framework, and we do not claim that it is universally better in every task. The DCN framework abstracts the core choices in SNN architectures, such as normalization and the polynomial basis. Different tasks might require different choices for these as shown in the empirics, and DCN architecture family encapsulates all, with a concrete derivation from the spectral filter view, and theoretical guarantees on asymptotic behaviour dependent on the choices. We believe R2L-DCN is a crucial architecture to have in the literature, for topological tasks on simplicial complexes that do not have a regular structure.

We motivate our architecture from R2L normalized operators, since it has the following properties:

  • Gives a normalized Hodge decomposition (not satisfied by up/down-degree based normalizations)
  • Bounds the eigenvalues aware of the local-geometry of the simplicial complex (which is not satisfied by 1/μmax⁡1/\mu_{\max})

Depending on the nature of the task, different normalizations can yield the same performance, which is expected. One of our main contributions is that how such architectures come from a spectral formulation, and what a polynomial basis should satisfy in order to be parametrized with block-diagonal and block-anti-diagonal matrices. This theoretical contributions opens up pragmatic avenues in SNN research, which can give insights on what models to use. Further discussions of the ablation and the choice of normalization is given in Appendix D.3.

We sincerely thank the reviewer again for their detailed feedback and constructive comments. We believe that we have clarified the novelty of our work and the depth study, and addressed their concerns with our experiments on filter transferability and ablations. We kindly ask the reviewer to reconsider their assessment in the light of our revisions. We are available in the case of any further questions.

AuthorsOpenReview ↗

Official Comment by Authors

Dear reviewer,

As the discussion period ends November 18 (AoE), we would love to hear if the additional experiments demonstrating filter transferability, ablations for isolating normalization, and clarifications of our novelty have resolved your concerns. In the meantime, we are available to discuss if there are any questions.

Reviewer_5ieiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper introduce Dirac convolutions, spectral filters on simplicial complexes defined separately for each coupled block D_I of the Dirac operator, acting on C_i-1 +o C_i with matrix valued frequency responses.

  1. polynomial from with block parity (Prop 3.4) : Any such filter is a matrix coefficient polynomial in D_[i] of the Dirac operator. Chirality forces even order terms to be block-diagonal (rank-wise, via Hodge Laplacians) and odd-order terms to be block anti-diagonal (inter-rank, via (co)boundaries).

  2. Chebyshev factorization (Cor. 3.5) : With Chebyshev basis, the odd terms factor as boundary maps times third-kind Chebyshev polynomials of 2L^~ - I. The instantiated model, R2L-DCN, uses the root-to-leaf (R2L) normalized operators \cite Viganò et al. (2026), whose spectra lie in [-1,1], with polynomial degree R_i = 2. SCCNN, SCNN are recovered as special cases.

  3. Asymptotic analysis of the linear inter-rank term : In continuous time with skew-symmetric weights, the norm and Dirichlet energy are preserved (Prop. 4.1). In all other analyzed cases, including forward-Euler with either symmetric or skew symmetric weights (Prop. 4.2), the Rayleigh quotient converges to mu_max at explicit rates. The authors argue that normalization slows this convergence.

Critical strengths

Clean correct explain formulation about SNN. Prop. 3.4 cleanly links matrix-valued spectral filters on each D_i to polynomial filters, and its chirality argument explains why even and odd terms separate into rank-wise and inter-rank processing. That separation is usually imposed by hand in SCCNN-style architectures. Cor. 3.5’s third-kind Chebyshev factorization of the odd terms is a neat, practically usable identity. Showing that SCCNN, SCNN and Ebli et al. are parameter-restricted special cases gives the community a useful common vocabulary, and the framework is a natural place to study normalization and basis choices.

A well-motivated case for normalized higher-order operators. Existing SNNs use unnormalized Hodge Laplacians whose spectral scale grows with local structure, so the paper addresses a real and under-discussed issue. The book-complex example makes the scale-mismatch problem concrete. The depth study (Fig. 4) is the most striking empirical result: SCCNN degrades with depth while R2L-DCN stays flat, which is consistent with the paper’s thesis.

Critical weaknesses
  1. The theory does not support the stability and depth claims made in the abstract and introduction. The abstract says linear DCNs “can preserve both the Dirichlet energy and the norm” and that normalization “improves stability, allowing deep architectures.”

  2. The empirical evidence does not isolate the contribution, and most margins are within noise. The paper’s central claim is that R2L normalization (with a Chebyshev basis) is what helps. However, there is no controlled ablation that holds the architecture, polynomial degree and tuning budget fixed while varying only the operator. The missing comparisons are:

  • unnormalized D;
  • D rescaled by lambda_max;
  • Schaub et al. (2020)’s normalized 1-Laplacian for the edge-flow tasks;
  • monomial vs. Chebyshev basis.
Questions for authors
  1. Can you add a controlled ablation within the same DCN architecture and tuning budget? The variants would be: (a) unnormalized D with a monomial basis, (b) D/lambda_max(D) with Chebyshev, (c) Schaub et al. (2020) normalization on edge tasks, and (d) R2L. This is the comparison that would show whether R2L normalization specifically drives the gains

  2. Is there any experiment that tests transfer across complexes, such as training on the regular ocean mesh and evaluating on the irregularized ones, or training and testing on meshes of different resolution? Contribution 3 claims transferability is demonstrated.

3 comments
AuthorsOpenReview ↗

Official Comment by Authors - Part 1:

We sincerely thank the reviewer for the valuable and constructive comments, especially finding our contributions crucial for unifying different architectures and practically relevant with our discussions of spectral scale, polynomial basis choices, and the depth study. We address the reviewer's concerns below:

Filter Transferability Experiments

Q2) Is there any experiment that tests transfer across complexes, such as training on the regular ocean mesh and evaluating on the irregularized ones, or training and testing on meshes of di"erent resolution? Contribution 3 claims transferability is demonstrated

We thank the reviewer for the suggestion. To answer this, we design an edge-classification task, where the goal is to detect which edges are corrupted via ±1\pm1 in the existence of a gradient signal coming from the vertices, where the filters are trained on a small (60 vertices) simplicial complex and transferred to larger complexes. We also incorporate the random-walk (rw) variant of R2L upon suggestions of other reviewers and show the results in the tables below.

Table A: Balanced accuracy (%) in corrupted-edge detection task for filters of order K.

OperatorKn=60n=240Irreg. n=5200
Unnorm.189.1 ± 1.982.2 ± 0.277.5 ± 0.0
395.4 ± 0.859.8 ± 0.768.2 ± 0.4
1/μ_max188.8 ± 1.998.9 ± 0.165.4 ± 0.2
395.8 ± 0.799.9 ± 0.081.5 ± 1.3
R2L187.0 ± 1.296.6 ± 0.196.5 ± 0.0
388.4 ± 1.095.4 ± 0.295.9 ± 0.0
R2L(rw)194.6 ± 0.699.9 ± 0.099.7 ± 0.0
395.9 ± 0.799.9 ± 0.099.8 ± 0.0

Table B: False positive rate (FPR) and false negative rate (FNR) for the filters. Lower is better.

OperatorKn=240 FPRn=240 FNRIrreg. n=5200 FPRIrreg. n=5200 FNR
Unnorm.135.7 ± 0.40.0 ± 0.044.9 ± 0.10.2 ± 0.0
379.5 ± 1.31.0 ± 0.063.1 ± 0.70.6 ± 0.0
1/μ_\max11.1 ± 0.11.2 ± 0.30.3 ± 0.168.8 ± 0.3
30.3 ± 0.00.0 ± 0.00.0 ± 0.037.0 ± 2.7
R2L12.3 ± 0.14.6 ± 0.22.6 ± 0.04.4 ± 0.0
38.5 ± 0.30.8 ± 0.07.2 ± 0.01.1 ± 0.0
R2L(rw)10.1 ± 0.00.1 ± 0.00.2 ± 0.00.4 ± 0.0
30.2 ± 0.00.0 ± 0.00.3 ± 0.00.0 ± 0.0

We see that R2L normalization outperforms the unnormalized filters for transferability in this task. This happens as an unnormalized filter extrapolates the frequency response to unseen eigenvalues, where it disproportionately scales up the outputs compared to the learned threshold. R2L also outperforms the 1/\mu_\max normalization when the underlying simplicial complex is irregular. This is because the irregularity is caused by a dense, local cluster, and 1/\mu_\max disproportionately shrinks the filter outputs, causing the outputs to remain below the threshold. This is visible in Table B. In ocean drifters dataset, the frequency of interest is 0, i.e. the harmonic component, which unnormalized operators keep consistent across complexes.

We added Section 5.3 and Appendix D.2 in the paper including further details of the problem and presentation of the results.

AuthorsOpenReview ↗

Official Comment by Authors - Part 2:

Ablations

Can you add a controlled ablation within the same DCN architecture and tuning budget?... 

We thank the reviewer for mentioning further comparisons to isolate normalization. In this direction, we now add ablations on the Ocean Drifters dataset with 1/μmax⁡1/\mu_{\max}, 1/μmax⁡1/\mu_{\max}-Cheb with the Chebyshev polynomials, R2L-rw using the random walk Laplacian, R2L-Monomial with the monomial basis. We also would like to note that Schaub et al.'s normalization is a special case of the R2L normalization, with k=2k=2, while R2L works for all ranks.

Methodoceanocean (+150)ocean (+600)
R2L-Monomial57.7 ± 5.751.1 ± 9.8<ins>45.5 ± 4.6</ins>
1/μmax⁡\mu_{\max}51.5 ± 6.9<ins>50.1 ± 4.8</ins>42.3 ± 6.5
1/μmax⁡\mu_{\max}-Cheb53.9 ± 5.048.8 ± 6.948.7 ± 5.6
R2L-DCN<ins>56.7 ± 5.3</ins>46.8 ± 8.341.1 ± 4.0
R2L-Monomial (even)50.8 ± 5.441.6 ± 3.836.4 ± 4.5
R2L-rw (even)52.7 ± 4.543.9 ± 5.140.1 ± 6.6
1/μmax⁡\mu_{\max} (even)48.9 ± 7.245.1 ± 5.235.8 ± 8.7
1/μmax⁡\mu_{\max}-Cheb (even)53.6 ± 5.645.6 ± 7.343.4 ± 3.9
R2L-DCN (even)53.2 ± 5.448.0 ± 6.045.4 ± 5.0

In coupled models, R2L-Monomial beats all models in ocean and ocean(+150), performing reasonbly well in ocean(+600), where 1/μmax⁡\mu_{\max}-Cheb excels. In rank-wise models and R2L-DCN (even) performs the best in ocean(+150) and ocean(+600), and 1/μmax⁡\mu_{\max}-Cheb (even) performs best in ocean. Referring to the discussion in the previous part, the similar performance of the normalized models can be explained through the fact that the task is highly dependent on the harmonic frequency.

We also would like to clarify that R2L-DCN is an instantiation of our framework, and we do not claim that it is universally better in every task. The DCN framework abstracts the core choices in SNN architectures, such as normalization and the polynomial basis. Different tasks might require different choices for these as shown in the empirics, and DCN architecture family encapsulates all, with a concrete derivation from the spectral filter view, and theoretical guarantees on asymptotic behaviour dependent on the choices.

We motivate our architecture from R2L normalized operators, since it has the following properties:

  • Gives a normalized Hodge decomposition (not satisfied by up/down-degree based normalizations)
  • Bounds the eigenvalues aware of the local-geometry of the simplicial complex (which is not satisfied by 1/μmax⁡1/\mu_{\max})

Depending on the nature of the task, different normalizations can yield the same performance, which is expected. One of our main contributions is that how such architectures come from a spectral formulation, and what a polynomial basis should satisfy in order to be parametrized with block-diagonal and block-anti-diagonal matrices. This theoretical contributions opens up pragmatic avenues in SNN research, which can give insights on what models to use.

Stability

The theory does not support the stability and depth claims... linear DCNs “can preserve both the Dirichlet energy and the norm"... normalization “improves stability, allowing deep architectures.”

Regarding preservation of Dirichlet energy and the norm, as our abstract specifies, we only make the claims for continuous and linear DCNs, which is the setting of proposition 4.1, showing that continuous and linear DCNs can satisfy energy and norm-preservation if their weight matrix is skew-symmetric.

The stability claim for the forward-Euler discretization specifically states they behave 'more stable' compared to their unnormalized counterparts. What we meant by that is, for the (skew-)symmetric weights, the norm of the features grows by at most (1+τ2θmax2μmax)(1+\tau^2\theta_{max}^2\mu_{max}) at each step. When the leading term and when τ\tau and θmax\theta_{max} are fixed, normalization ensures ∣μmax∣≤1\vert\mu_{max}\vert \le 1, thereby providing a uniform bound on the feature norm across all simplicial complexes, while unnormalized complexes can have arbitrarily large eigenvalues (e.g. book complexes). We thank the reviewer for bringing this up, and agree that a better term to be used here is to say 'bound the growth of the norm uniformly across all simplicial complexes', and we have revised the abstract, contributions, and section 4 accordingly, and added more discussion on how the norm and convergence are uniformly bounded across all SCs in Appendix C.1.

Once again we thank the reviewer for their engagement and careful review of our research. We appreciate the constructive suggestions and hope that we were able to answer the concerns about the examples of filter transferability, ablations and clarification of the stability contribution.

AuthorsOpenReview ↗

Official Comment by Authors

Dear reviewer,

As the discussion period ends November 18 (AoE), we would love to hear if the additional experiments demonstrating filter transferability (Q2), ablations isolating normalization (W2 / Q1) and the discussions on stability (W1) have resolved your concerns. We are happy to answer any questions in the meantime.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment & Thanks

Dear reviewers and AC,

We thank the reviewers for their detailed feedback and constructive suggestions, mainly on filter transferability, clarifications of novelty, stability, and ablations, which helped us to improve the paper. Although we could not have a discussion with 2 of the 3 reviewers, we kindly ask them to reconsider their assessment in the light of the additions we made to the paper during the discussion period, addressing each one of the concerns. The paper is updated now, with the changes highlighted in blue. We believe we addressed all of the concerns of the reviewers, and the suggestions overall strengthened the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.